Baidu’s ERNIE 5.5 Flash Base is a pretrained base model and Moonshot’s Kimi K3.1 Pro is a post-trained multimodal model. Both publish weights under custom licences, and Baidu’s excludes the EU, the UK and South Korea.
Two releases with different purposes
Baidu published ERNIE 5.5 Flash Base on November 5, 2026, according to Baidu’s release blog, as a text-only mixture-of-experts (MoE) model with 280 billion total parameters. The weights come in an FP8 repository and a separate BF16 repository. Moonshot AI made Kimi K3.1 Pro available on its products and API at launch and said in its launch blog that the full weights would follow by November 4, 2026. They are now on the Kimi K3.1 Pro model card, a text-and-image MoE model with 2.8 trillion total parameters.
Anyone can download either set of weights without an access request, and each has its own licence. That is where the two diverge most: ERNIE’s grants rights only outside the European Union, the United Kingdom and South Korea, while Kimi’s file states no geographic limit. The purposes differ as well, because one is a starting point for training and the other is a finished model.
| Specification | ERNIE 5.5 Flash Base | Kimi K3.1 Pro |
|---|---|---|
| Developer | Baidu | Moonshot AI |
| Architecture | MoE | MoE |
| Total parameters | 280B | 2.8T |
| Active parameters per token | 14B | 104B |
| Modality | Text | Text and image |
| Context length | 1,048,576 tokens in fine-tuning and inference, trained at up to 128K | 1,048,576 tokens |
| Checkpoint stage | End of pre-training | Post-trained |
| Weights size | About 288 GB (FP8) or 573 GB (BF16), in two repositories | About 1.56 TB (MXFP4), in one repository |
| Licence | Custom agreement, excludes the EU, the UK and South Korea | Custom licence, no geographic limit in the file |
Sources: ERNIE model card and Kimi model card for the parameters, modality, context and training stage. The sizes are the safetensors totals Hugging Face lists for the FP8, BF16 and Kimi repositories, read on November 12, 2026. The licences are covered below.
ERNIE 5.5 Flash Base
Baidu says the model was pretrained from scratch on 33.5 trillion tokens with a training context of up to 128K, and that fine-tuning and inference support 1M tokens (Baidu’s blog). The model card states 1,048,576 tokens as the figure for fine-tuning and inference.
The model card describes the weights as the checkpoint “at the end of pre-training, before any mid- or post-training”. It was published before mid-training and post-training, so it is a starting point for domain adaptation, instruction tuning and reinforcement-learning research, not a finished conversational agent. Baidu’s vLLM-ERNIE guide lists the base model as a completions model and lists a chat, reasoning and tool-calling ERNIE-5.5-Flash as “Coming soon”. Baidu counts the 14B active parameters without the embedding lookup, and its totals exclude 0.36B parameters of the ECHO layer, a note on the model card.
Kimi K3.1 Pro
Moonshot says Kimi K3.1 Pro was post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains, and that it brings gains in coding, 3D creation, financial research and factual accuracy (model card). The card lists MXFP4 weights with MXFP8 activations, trained with quantization-aware training. It names vLLM, SGLang and TokenSpeed for serving, and lists text and image input.
Moonshot’s blog says it also released Kimi K3.1 Pro Highspeed, which it describes as delivering two to three times the speed of K3.1 Pro. The blog also describes a security step: Moonshot says that for the open-weight release it identified and removed cyber-related data from mid-training and from its reinforcement-learning environments. On its in-house xdev-bench-hard, Moonshot reports that the final model solves 81.0% of problems at pass@3, against 90.5% for a checkpoint without those measures. That is Moonshot’s own benchmark and measurement.
Licence terms
Both repositories carry the Hugging Face licence label “other”, which marks a custom text. TechiesJournal does not call either one open source. This section summarises the files as read on November 12, 2026. It is not legal advice, and readers should read the licence for their own situation, and ask counsel where the territory, revenue or service terms may apply.
Where each licence allows use
The ERNIE 5.5 Flash Base Model License Agreement defines its Territory as the world other than the European Union, the United Kingdom and South Korea (section 1.11). The agreement grants Baidu’s rights “solely within the Territory” (section 2). Section 5.3 says the weights, derivative models and their outputs must not be used, reproduced, modified, distributed or displayed outside it, and the acceptable use policy in Exhibit A repeats the point as its first item. Under Baidu’s licence, use inside the European Union, the United Kingdom or South Korea is outside the rights granted. Anyone there should read the agreement and take legal advice. The agreement is governed by Hong Kong law, with Hong Kong courts (sections 9.1 and 9.2).
The Kimi K3.1-Pro License has no geographic restriction. It asks that use comply with applicable laws and regulations (section 1).
Commercial use and other conditions
| Term | ERNIE 5.5 Flash Base | Kimi K3.1 Pro |
|---|---|---|
| Non-commercial use | Free of a model licence charge (section 4.1) | Free of charge, subject to the conditions in sections 1 to 4 |
| Commercial use | Free of a model licence charge up to US$50 million of annual revenue from related products or services, with a one-time use notice to Baidu (sections 3.2, 4.2) | No charge in general. A Model as a Service business with more than US$20 million of revenue over any 12 months needs a separate agreement with Moonshot (section 2) |
| Above the threshold | Separate prior written authorisation from Baidu, and commercial use without it must be suspended (section 4.3) | Interface credit “Kimi K3.1-Pro” for products above 100 million monthly active users or US$20 million monthly revenue (section 3) |
| Attribution | “Powered by ERNIE” on any commercial product or service (section 3.2) | Keep the copyright and permission notice in all copies (section 1) |
| Use of outputs | Outputs may not be used to improve any other AI model, and distillation is treated as a derivative model (sections 5.5, 1.5) | The file states no restriction on outputs or distillation |
| Exemptions | The Territory limit applies to all use (sections 2, 5.3) | Internal use, and use through Moonshot’s official products or certified inference partners, are outside sections 2 and 3 (section 4) |
| Other limits | Acceptable use policy, indemnity and patent-suit termination (sections 6.6, Exhibit A) | Warranty disclaimer (section 5) |
Sources: ERNIE licence file and Kimi licence file, read on November 12, 2026.
Two details in the ERNIE text matter for planning. Revenue counts the whole related product, not only the share attributable to ERNIE, and it aggregates every entity under common control (section 1.7). Baidu can also ask an operator whose revenue is at or above US$50 million for a compliance statement, according to section 4.5 of the agreement. Exhibit A, which the agreement shows as last revised on November 6, 2026, bars among other things military uses, high-stakes automated decisions in areas such as healthcare, credit and employment, and publishing generated material without clear identification as machine-generated. For Kimi, “Model as a Service” means giving a third party access to inference or fine-tuning, for example through an API, with meaningful control over inputs, parameters or training data. Terms can change after publication, so recheck the files before relying on them. TechiesJournal’s earlier article on DeepSeek, GLM and Kimi open-weight models covers how to think about this kind of custom licence.
What the evidence shows
Almost all of the performance evidence is from the two vendors. The table separates who measured what.
| Source | What was measured | Result |
|---|---|---|
| Moonshot, in the model card | Terminal-Bench 4.0, Kimi run with the Claude Code harness | Kimi K3.1 Pro 65.7, Claude Opus 5.5 66.4, GPT-6 Astra 58.2 |
| Vals AI | Terminal-Bench 4.0 | Kimi K3.1 Pro 63.64%, fourth of 63 models |
| Vals AI | Vals Index, finance, coding, legal and tax tasks | Kimi K3.1 Pro 62.52%, rank 10. Kimi K3 52.95%, rank 28 |
| Baidu, in the model card | 30 few-shot benchmark rows against three other base models | Baidu marks ERNIE best in 24 rows and a rival best in 6 |
Sources: Kimi model card, Vals AI’s Kimi K3.1 Pro page, the Vals Index leaderboard, updated November 9, 2026, and the ERNIE model card.
For Kimi, Vals AI is the one independent measurement found. Its Vals Index scores Kimi K3.1 Pro at 62.52% against 52.95% for Kimi K3, a gain of 9.57 points by our arithmetic, which is consistent with Moonshot’s account of what extra post-training bought. Several caveats apply:
- Vals tests different tasks from Moonshot’s tables, and nine models rank above Kimi K3.1 Pro on the index, among them Claude Fable 5.2 at 71.31% and, in the same table, Claude Opus 5.5 (70.14%) and GPT-6 Astra (65.02%), two of the models Moonshot compares against.
- Moonshot’s Terminal-Bench figures are a mix of harnesses. The card says the comparison scores, apart from DeepSeek’s, are the maximum values reported in model blogs or on the official leaderboard.
- Vals AI’s note on its Kimi K3.1 Pro evaluation is dated October 24, 2026, before the weights had been released, so the score most likely describes the hosted model and not a copy you would run yourself. That is our inference.
- Vals AI prints Kimi K3’s score as 52.95% on its leaderboard and 52.96% in its note on Kimi K3.1 Pro.
For ERNIE 5.5 Flash Base, no independent measurement of the checkpoint was found. Baidu’s table compares it with DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base, run by Baidu with the same pipeline, prompts and few-shot settings. The rivals lead on six rows: ChineseSimpleQA, BBH with chain of thought, INCLUDE-lite, FullStackBench, HELMET-Rerank and InfiniteBench. Base-model few-shot scores are not comparable with chat-model leaderboards such as Vals AI’s. TechiesJournal’s guide to how to read AI benchmark claims covers what to ask of numbers like these, and its article on Tencent’s Hy4 reads the same Vals leaderboard.
Hardware and weight size
The model cards describe sparse models in which only some experts run per token, but every expert must still be stored, so the files stay large. The table puts the sizes beside what each vendor or serving project says about the smallest setups.
| Weights | Size | Setup stated by the vendor or serving project |
|---|---|---|
| ERNIE FP8 | About 288 GB | Baidu’s guide: 2 GPUs of 288 GB, 4 of 192 GB or 8 of 80 GB |
| ERNIE BF16 | About 573 GB | Baidu’s guide: 4 GPUs of 288 GB, 8 of 192 GB or two nodes of 8 GPUs of 80 GB |
| Kimi K3.1 Pro MXFP4 | About 1.56 TB | The vLLM recipe: tensor parallel 8 on one B300 node or two GB300 trays, or on one MI355X node |
Sources: Baidu vLLM-ERNIE guide for the ERNIE rows, which it describes as the smallest GPU counts that hold the weights, and the vLLM recipe for Kimi K3.1 Pro, which the model card links and which is the vLLM project’s, not Moonshot’s. The sizes are from Hugging Face.
These counts only hold the weights, and nothing here is a measured throughput. Baidu’s guide runs its example servers at a 131,072-token context and says 80 GB GPUs fit up to 262,144 tokens but not 1M, so the full context needs more memory. The vLLM recipe says throughput was not tuned. TechiesJournal’s perspective on hosted AI versus self-hosted LLMs weighs when running a model yourself is worth that cost.
What developers should know
Both sets of weights can be downloaded today, but neither is a desktop model. By the sizes and the Baidu and vLLM guides cited above, ERNIE needs 288 GB to 573 GB of storage and at least two data-centre GPUs, and Kimi K3.1 Pro needs about 1.56 TB, with the vLLM recipe’s baseline an eight-GPU group. ERNIE 5.5 Flash Base is a base checkpoint for teams that will train it further, and only inside Baidu’s Territory. Kimi K3.1 Pro is a post-trained model for applications and agentic tasks, and its licence conditions apply mainly to Model as a Service businesses and very large products. Test either on your own tasks, and recheck both repositories and licence files before relying on them.
References
- Baidu: ERNIE 5.5 Flash Base release blog, November 5, 2026
- ERNIE-5.5-Flash-Base model card (FP8)
- ERNIE-5.5-Flash-Base-BF16 model
- ERNIE 5.5 Flash Base Model License Agreement
- Baidu vLLM-ERNIE on GitHub
- Moonshot: Kimi K3.1 Pro launch blog
- Kimi-K3.1-Pro model card
- Kimi K3.1-Pro License
- vLLM recipe for Kimi K3.1 Pro
- Vals AI: Kimi K3.1 Pro model page
- Vals AI: Vals Index leaderboard
Sources were read on November 12, 2026. Specifications, licence terms, benchmark results and availability are Baidu’s, Moonshot’s, vLLM’s and Vals AI’s, and can change.